The Helios orchestrator is designed to manage virtual machines on preemptible instances, which offer cost savings of up to 70–80% compared to standard pricing. The system architecture follows a hyperconverged infrastructure model, integrating storage, compute, and networking within each node to eliminate the need for complex external topologies or specialized enterprise hardware.
To ensure high availability despite the potential for sudden host termination, Helios uses a three-layer control plane. It employs the Raft consensus algorithm for individual resource shards, a Gossip protocol to maintain a real-time cluster map, and a Reconciler to align the actual state of physical resources with the desired configuration stored in Raft.
The system enables seamless live migration of virtual machines within a 120-second window before a host is shut down. By utilizing parent subscriptions, the virtual machine shard acts as a non-voting Raft learner, allowing it to receive immediate state updates from storage and network shards without relying on external timers or queues.
Helios uses DRBD for storage replication and Cloud Hypervisor for managing virtual machine processes. The architecture avoids centralized metadata stores, distributing the control plane across all nodes to prevent bottlenecks and ensure that the data plane does not interfere with cluster management tasks.
